Inspiration
Many people want to use AI on sensitive data like legal documents, business plans, and financial records, but they fear privacy risks and cloud data leaks. This inspired us to build Undercover AI, a system that works fully offline and keeps all data private on the device.
What it does
Undercover AI analyzes business plans, legal documents, bank statements, and proprietary code — without sending any data to the cloud. It provides summaries, risk detection, financial insights, and contract analysis with low latency and zero privacy risk.
How we built it
Undercover AI using on‑device Small Language Models (SLMs) optimized with quantization to run on mobile phones. We use the RunAnywhere SDK, offline OCR, and secure local storage to ensure fast, private, and offline AI processing and we use Llama-3-3B (Quantized) and DeepSeek-R1-Distill (Quantized).
Challenges we ran into
1.Running AI models on mobile with limited memory 2.Reducing model size while keeping good accuracy 3.Improving speed without draining battery 4.Designing a fully offline workflow with strong privacy
What we learned
We learned how to: 1.Optimize AI models for mobile devices 2.Build privacy‑preserving AI systems 3.Balance performance, accuracy, and efficiency 4.Design real‑world solutions for sensitive data problems
What's next for Undercover AI
1.Improve model accuracy and speed 2.Add more document types and industries 3.Support enterprise‑level offline deployment 4.Expand into legal, finance, healthcare, and compliance use cases
Log in or sign up for Devpost to join the conversation.